feat: update voices

generate_voices.py

@@ -14,15 +14,15 @@
 model = Supervoice(None, None, encodec_model, None, tokenizer)
 
 # Find all wav files in the voices directory
-wav_files = list(Path('voices').glob('*.wav'))
+wav_files = list(Path('voices').glob('*.flac'))
 wav_files = [f.stem for f in wav_files]
 
 # Generate voices
 for id in wav_files:
     print(f"Processing {id}")
     with open("./voices/" + id + ".txt", 'r') as f:
         text = f.read().strip()
-    created_voice = model.create_voice(audio = "./voices/" + id + ".wav", text = text)
+    created_voice = model.create_voice(audio = "./voices/" + id + ".flac", text = text)
     torch.save(created_voice, f"./voices/{id}.pt")
 
 # Generate index file

supervoice_valle/model.py

@@ -41,23 +41,17 @@ def create_voice(self, audio, text):
 
         # Prepare text
         text = self._normalize_text(text)
-
-        # Tokenize text
-        text_tokens = self.tokenizer.encode(text)
 
         # Return
         return {
             "audio_tokens": audio_tokens,
-            "text_tokens": text_tokens,
+            "text": text,
         }
 
     @torch.inference_mode()
     def synthesize(self, voice, text, top_k = None, top_p = 0.2):
         device = self._device()
 
-        # Prepare text
-        text = self._normalize_text(text)
-
         # Prepare voice
         if type(voice) is str:
 
@@ -73,7 +67,7 @@ def synthesize(self, voice, text, top_k = None, top_p = 0.2):
             voice = torch.load(voice_file, map_location = "cpu")
 
         # Tokenize text
-        text_tokens = torch.cat([voice["text_tokens"].to(device), self.tokenizer.encode(text).to(device)])
+        text_tokens = self.tokenizer.encode(self._normalize_text(voice["text"]) + " " + self._normalize_text(text)).to(device)
 
         # Audio tokens
         audio_tokens = voice["audio_tokens"].to(device)        

supervoice_valle/voices_gen.py

@@ -1 +1 @@
-available_voices = ['voice_3', 'voice_1', 'voice_2']
+available_voices = ['emo_2', 'voice_1', 'emo_1', 'voice_2']

voices/emo_1.txt

@@ -0,0 +1 @@
+Ok, but one word!

voices/emo_2.txt

@@ -0,0 +1 @@
+They are the most beautiful a whole court!

voices/voice_1.txt

@@ -1 +1 @@
-We had not stopped to study the Indian character.
+No will powerful enougth to find a way to carry out my wishes.

voices/voice_2.txt

@@ -1 +1 @@
-A man approached through the heavy gloom.
+And even if both of them has survived the terror.